Model inference low-level optimization, hardware affinity and large-scale distributed training.